pku-mocca/DexterHand

DexterCap: An Affordable and Automated System for Capturing Dexterous Hand-Object Manipulation

Data Visualization

For a quick overview and inspection of the dataset, please visit the DexterHand Interactive Visualization page.

Data Description

The dataset is stored in multiple .npz files, each of which is a mocap session.

Each .npz file contains the following keys and structure:

Metadata (data['metadata'])

• description_string: A string describing the dataset.
• mocap_session_name (str): Name of the motion capture session, e.g. "Cuboid_00".
• index_init_frame (list[int]): List of frame indices where new action sequences start, used to segment the data into separate action sequences. For example, [0, 10] means the first sequence starts at frame 0, and a new sequence starts at frame 10.
• invalid_point_value (list[float]): If the object is not visible in a frame, its translation will be set to invalid_point_value , e.g. [-1000, -1000, -1000].
• object_class (str): The class/type of the object, e.g. "Cuboid".
• object_size (list[float]): The size of the object. For a cuboid, this is [length, width, height], e.g. [0.15, 0.04, 0.04].
• fps (int): The frames per second of the data, e.g. 60.

Hand Data (data['hand_translations'] and similar keys)

• hand_translations (np.ndarray, shape (N, 3)): The 3D translation of the hand for each frame.
• hand_orientations_axis_angle (np.ndarray, shape (N, 3)): The 3D axis-angle rotation of the hand for each frame.
• hand_poses (np.ndarray, shape (N, 45)): The pose parameters of the hand for each frame, 15 joints with axis-angle representation.
• hand_shapes (np.ndarray, shape (N, 10)): The shape parameters of the hand for each frame.

Object Data (data['object_translations'] and similar keys)

• object_translations (np.ndarray, shape (N, 3)): The 3D translation of the object for each frame.
  If the object is not present in a frame, the value will be invalid_point_value.
• object_orientations_quat_xyzw (np.ndarray, shape (N, 4)): The orientation of the object in quaternion format [x, y, z, w] for each frame.